Since the first demo of ChatGPT was released in November 2022, the pace and capability of AI-first companies has been truly mind-boggling. The rapid development of AI capability has led to an explosion in the number of companies seeking funding. Funding raised by Generative AI businesses quadrupled between 2022 and 2023 ($612m to $2.3bn) and that trend is continuing in 2024. Since the start of the year we have had 493 companies pitch to us at Ada Ventures. That is amongst the busiest quarter for dealflow since we launched the fund. The vast majority of these companies have been AI-first. What I mean by AI-first is that these companies are building with AI as a key part of the value proposition. Some of these are Generative AI companies, some have been built on top of LLMs and others are using AI in other ways — for example to process and analyse large volumes of data. It’s not just Ada that have been busy with AI-first companies. Martin Mignot, Partner at Index stated that January & February 2024 were the busiest months in his 14 years at Index. 50% of the investments Index made in these two months were into Generative AI companies.

Despite all the noise in the early stage tech ecosystem — Generative AI is still a relatively nascent market in the wider economy and we believe there’s a lot of value to be built and captured. At the end of 2023, Gen AI represented less than 1% of enterprise cloud spend.

This post covers:

1. Why we at Ada are interested in AI-first approaches and how they relate to Inclusive Alpha®
2. How we evaluate AI-first companies at pre-seed and where we are focused
3. Some of the impact risks of AI-first companies that we are wrestling with

‍

Part I: Why we’re particularly interested in AI-first companies

At Ada Ventures we are enthusiastic about the developments in AI-first companies for different reasons than many others. We believe in Inclusive Alpha® — that investing in an inclusive way will drive best in class returns for our investors. The rapid developments in AI are interesting to us for these key reasons:

Who we invest in: Advances in Generative AI meaningfully lower the fundraising and academic barriers to entry to found a software company. We believe this will lead to an increase in the number of underrepresented founders able to build companies. In particular we are already seeing this play out in founders building Application layer* (defined below) companies on top of open sourced models (Llama, Hugging Face). Because founders don’t have to build the whole stack themselves they are also able to get to revenue and product market fit faster, using co-pilots like Google’s Gemini Code Assist and GitHub co-pilot to “boost” engineer productivity. We also believe this should mean that they are not as disadvantaged by the fact that underrepresented founders raise smaller rounds.

How we invest: Generative AI is also increasing the ability for Ada Ventures ourselves to reach founders in places where others aren’t looking. We are already working with scouts and angels to source 10x more all-female teams and 6x more all-black teams at the top of our pipeline. We are also sourcing via social media platforms in order to seek out founders who don’t have a ‘warm introduction’ to a VC fund through their networks. We’re now using GenAI to speed up market research, enabling us to have a prepared mind and move quickly to hone in on the most impact areas of an investment case (more detail on this in Michael’s post here). Finally, GenAI is helpful for our stage focus — we invest at pre-seed and seed and we believe that founders will not need to raise as much capital as previously to reach product-market-fit.

What we invest in: We invest in companies solving hard problems in Healthy Ageing, Climate Equity and Economic Empowerment. We believe that breakthroughs in AI could help us to find new approaches to tackle some of the persistent problems in these spaces that haven’t been cracked yet. There are three types of company that we are most interested in:

• Unique & underexplored datasets: Companies accessing previously untapped datasets and using AI to analyse them. For example, Juno Bio, an Ada Ventures portfolio company is using machine learning to examine the Vaginal Microbiome, a dataset which is so big that it was impossible to process and find patterns without AI. With Juno’s technology they have discovered links in the vaginal microbiome to endometriosis, pre-term birth and cancer. Similarly, Ada Portfolio company Materials Nexus is modelling the periodic table of elements to discover new materials. Advances in computing and machine learning allow them to predict the properties of different compounds before they are tested and developed in a laboratory. This allows the company to whittle down candidates for development from 5 million potential compounds to 10 thereby bypassing the slow process of discovery by trial and error.
• Increasing access: Companies leveraging AI to meaningfully lower the price for an important product or service, broadening access to a wider group. With advances in generative AI and inc reading efficiencies we believe the average person and ultimately lower income groups will be able to afford to access things like wealth management, or legal advice, or consultant led care that was previously prohibitively expensive. We believe that Jevons Paradox will apply so despite the pricing of these services reducing significantly, the overall market will expand. A good example of this is Do Not Pay and Valla here in the UK.
• Focusing on points of maximum friction: Companies using AI to tackle some of the most persistent and difficult challenges that underpin social issues — for example educational attainment, access to affordable housing, access to quality mental health support. We are meeting founders who are tackling the ‘points of maximum friction’ — for example, planning laws that make building new affordable housing difficult and expensive, or the complicated benefits system. Generative AI is enabling founders to address these points of friction by processing massive data sets to pinpoint opportunities more quickly.

‍

Part II: Evaluating AI-first Companies at Pre-seed

We first published our model for evaluating companies (A Seed Investing Framework) six years ago, and still apply a version of this framework today. We believe you need an evolved framework for Investing in AI-first companies at pre-seed and seed. I hope that it is a useful prompt to founders already working and those considering building in this space. This framework is meant to spark a discussion and I would love input here. What have we missed, why is this wrong? Please add your thoughts and feedback in the comments.

Very simply we see there being three ‘layers’ to the AI-first Company (software) stack. For each of these ‘layers’ there are different considerations that an investor has to weigh up — so we’ve mapped these on a graph with the X-axis being capital intensity and the y-axis being defensibility. Our hypothesis is that Infrastructure Layer companies are the most capital intensive but also more defensible, application layer companies are the least defensible but also the least capital intensive. For the Applied Infrastructure Layer, the companies are less capital intensive than building a foundational model, but still require hiring top developer talent and large engineering teams. They likely have a longer go to market (usually bottom up via developers) and therefore have a longer time to value. They get more deeply embedded into the stack and therefore switching costs are higher and are therefore more defensible.

We think that there are very valuable companies to be built at both the Applied Infrastructure and the Application Layers — but there are different considerations for founders and investors. Our Framework for what we’re looking for in Applied Infrastructure and Application Layer companies is below, alongside our ideal entry price and the key risks we are looking at at each layer.

‍

Infrastructure Layer — Eg. LLMs. Company examples: Mistral, Open AI, Ada Portfolio eg: N/A

At Ada Ventures, we are generally not investing in Infrastructure Layer companies. We don’t have deep enough pockets to fund the compute costs for Large Language Models that require massive GPU clusters, especially with sovereign funds and trillion dollar tech companies funding them. There was initially a view that these companies would be defensible due to their data advantages, but more recently we’ve seen similar performance from models trained on smaller datasets and also customers willingness to switch from one model to another. So whether these truly are more defensible than other types of AI-first companies is unclear. However — these are still very capital intensive to build and that does create a moat of sorts.

‍

Applied Infrastructure Layer (Eg. Cyber Tooling for LLMs, model supervision, fine tuning). Eg LangChain, Ada Portfolio eg. Occam AI (Occam AI sells an ‘embedded agents’ API that are designed to give existing enterprise software intelligent automation capabilities).

These companies are building core technology that adapts some aspect of the Infrastructure Layer to make it more usable or address a more specific use case. For example, a company building cyber security tooling which sits on top of an LLM which makes it more secure, in order that it can be usable by financial services companies. Applied Infrastructure companies typically go to market via developers and help developers ship more products more quickly and ensure that those products are performing correctly. The depth of focus on a technical breakthrough means that these companies tend to be more capital intensive since they tend to need bigger engineering teams. However, the prize could be bigger if they are successful.

Application Layer (Eg. Copilot for lawyers eg. Harvey). Ada Portfolio eg. Gizmo, PlannerPal.

Finally, the least capital intensive companies are the Application Layer companies. These can typically be built with a small team and companies can start to achieve revenue traction as companies and individuals are experimenting with new tools. Application Layer companies are those that are using GenAI, often building on top of closed or open-source models and building applications to solve end user problems, whether B2B or B2C users. This layer could be an attractive place to invest. We are particularly intrigued by vertical companies building customised solutions to their industry (for example PlannerPal in Wealth Management) and marketplaces (B2B or B2C) like Gizmo in education. With this layer, founders must be mindful of building a moat and increasing defensibility to build a true long term competitive advantage. Compelling ways that we have seen founders build defensibility in this layer are:

• Encouraging user generated content — by incentivising users to add their own proprietary data to the product, the product is personalised and enriched, making it stickier.
• Tackling a regulated industry which has specific domain specific requirements, meaning that generic LLMs find it harder to tackle.
• Being an early mover in the market and getting embedded within the workflow and used daily by customers.
• Building for a specific vertical so that the tools, language and approach can be tailored to suit this vertical.
• Network effects in the business model

This framework is meant to spark a discussion and I would love input. What have we missed? Please add your thoughts and feedback in the comments.

‍

Part III: Impact risks for AI-first companies

There are a range of risks (privacy, transparency, bias and more covered comprehensively in this piece) inherent to building with AI. Below are the key risks we are thinking most about when investing in AI-First companies at the pre-seed stage:

1. Automation and job loss. This is happening already and we believe it will increase, with the potential to increase overall inequality. The IMF estimates that 40% of global jobs will be affected by AI. In advanced economies, the figure is estimated to be 60%, with half benefitting from AI integration, enhancing productivity. For the other half, AI applications may execute key tasks currently performed by humans, which could lower labor demand, leading to lower wages and reduced hiring with a number of jobs. How we think about this at Ada Ventures is that we can’t stop this from happening. What we can do is (within our Economic Empowerment theme) — invest in areas that will upskill, reskill and retrain people that will lose white collar jobs. For example our portfolio company Greenworkx which upskills and reskill people towards Green Jobs for the climate transition. We will also invest in areas which will enable people to move away from very low skill and low paid, unsatisfying work and into higher paid work. This has the potential to exacerbate existing racial, socioeconomic and gender inequities that exist in our society.
2. Safeguarding vulnerable users. Tackling markets with end users that are vulnerable (groups with protected characteristics) requires founders to be incredibly thoughtful. There is a duty of care to vulnerable people who may implicitly trust AI — or not know that they are speaking to a machine at all. Safeguarding and ethics will be incredibly important in conversations with founders building products for these users.
3. Ethical, legally accessible and debiased datasets. Due to the speed with which the landscape is moving, datasets used to build AI-first companies have not always been held to a high standard for accuracy or diversity. The legal status of training data under ‘fair use’ is also an area which is currently ambiguous and disputed. For example, recently Ed Newton Rex wrote about Suno and their AI Generated music which bears a close relationship to copyrighted hit pop music tracks. I’m encouraged by the work Ed is doing with Fairly Trained and look forward to seeing more organisations like his tackling this problem.
4. Environmental impact of compute. At Ada we are investing in technologies that reduce the impact of the climate crisis or remove carbon. It’s therefore important that we weigh up any positive impacts made by these companies with the carbon emissions generated by the compute required to train the modes. A study from MIT suggested that generating an image using AI requires as much energy as charging your smartphone.
5. ‍

Will we at Ada Ventures still invest in non-AI-first companies?

Absolutely, yes. The low barriers to entry and limited defensibility of these AI companies make hard tech companies with genuine moats more attractive. We will still invest in companies that are not using AI as their starting point. However, we expect every company to be incorporating AI into their workflows to become more efficient and get to value faster.

We are actively investing and have closed four new investments so far in 2024, with another in completion. If you’re working on a breakthrough idea across Climate Equity, Economic Empowerment or Healthy Ageing, please get in touch.

If you have any builds on or questions about this piece, please let me know in the comments below.\

‍

Further reading

The Ethics of Advanced AI Assistants — Google DeepMind

Sell Work, Not Software — Sarah Tavel

The emerging vertical AI landscape — Cowboy Ventures

How to build defensible Generative AI Companies — Bola Adegbulu — former Ada Venture Partner

Verity Harding — AI Needs You

The AI50 — Visualised, Sequoia Capital

AI-First Companies — How to Compete and Win with Artificial Intelligence — Ash Fontana

The State of AI Report 2023 — Nathan Benaich, Air Street Capital

One More Thing — Ethan Mollick

The AI Workforce is Here: The Rise of a New Labor Market — Pete Flint & Anna Piñol

‍